SHAP Values

The sum of all feature attributions equals the difference between the model's prediction for a specific molecule and the average prediction across the dataset. This property, also known as efficiency, ensures the explanation is a faithful decomposition of the output. For a toxicity prediction model, if the base value is 0.3 and the prediction is 0.8, the SHAP values for all molecular fragments will sum to exactly 0.5.

A feature that is already missing from the input—such as a structural fragment not present in the molecule—is guaranteed an attribution of zero. This prevents the explanation from arbitrarily assigning importance to absent substructures. In molecular fingerprinting, bits representing absent circular substructures will not influence the SHAP explanation for that compound.

If a model is retrained so that a feature's marginal contribution increases or stays the same regardless of other features, the SHAP value for that feature will never decrease. This formal guarantee ensures that explanations track model behavior logically. If a retrained hERG cardiotoxicity model relies more heavily on a basic amine substructure, its SHAP value will reflect this increased dependence.

SHAP values are linearly additive across models. For an ensemble method like a random forest predicting logP, the SHAP value for a functional group is the average of its SHAP values across all individual trees. This property extends to explaining multi-task models, where the importance of a molecular feature for a combined ADMET score can be decomposed into its contributions to individual endpoints.

If two features contribute identically to the model's output in every possible coalition of other features, they receive identical SHAP values. In molecular property prediction, two topologically equivalent atoms in a symmetric molecule—such as the two methyl groups in para-xylene—will be assigned equal importance for any prediction, preserving chemical symmetry in the explanation.

Exact SHAP computation is exponential in the number of features, making it intractable for high-dimensional molecular representations. Kernel SHAP is a model-agnostic approximation that uses a weighted linear regression over sampled coalitions. For a molecule encoded with 2048-bit ECFP4 fingerprints, Kernel SHAP samples subsets of bits to estimate their contributions without evaluating all 2^2048 possible coalitions.

A gradient-based attribution method that satisfies the completeness axiom by accumulating gradients along a path from a baseline input to the actual input. Unlike SHAP, it is designed specifically for differentiable models like neural networks. For molecular property prediction, it can attribute a toxicity score to individual atoms by integrating gradients through a message-passing neural network.

A critical distinction in explainability:

• Feature Importance: Global, model-level metrics (e.g., permutation importance, Gini impurity) that rank features by their average contribution across all predictions.
• Feature Attribution: Local, instance-level explanations (e.g., SHAP values) that decompose a single prediction into feature contributions. In drug discovery, global importance identifies which molecular descriptors generally influence logP, while SHAP explains why a specific molecule received a high clearance prediction.

An adaptation of SHAP for deep neural networks that combines SHAP values with DeepLIFT's linear composition rule. DeepSHAP approximates SHAP values by backpropagating contributions through the network layers. For molecular graph neural networks, it can assign importance scores to individual atoms and bonds, revealing which functional groups drive a binding affinity prediction.

The original model-agnostic SHAP implementation that estimates Shapley values using a weighted linear regression over a sampled coalition space. While guaranteed to converge to true SHAP values, KernelSHAP is computationally expensive for high-dimensional molecular fingerprints (e.g., 2048-bit ECFP vectors). It serves as a fallback when model-specific methods like TreeSHAP or DeepSHAP are unavailable.